Conventionally, translation apparatuses for machine-translating a sentence input by a user to a different language have been known.
Main approaches in machine translations include rule-based approach and example-based approach. Outlines of these two approaches will be briefly described.
In the rule-based approach, translation is done using a word dictionary and rules and processes including grammar rules, semantic processing and generative rules.
In the rule-based approach, in principle, any sentence can be an object of translation as long as the word dictionary stores sufficient number of words. Even if an input sentence includes a word not in the word dictionary (unknown word), a method to handle such an unknown word is established in each system adopting such an approach.
On the other hand, the rule-based approach is not free from analytical error. If there is an error in the analysis, correct translation does not result. Even if the result of analysis is correct, the user may feel the generated translation awkward, as it is synthesized from the rules and dictionary.
In the example-based approach, translation is done using, for example, parallel translation examples and matching means for finding a match between an input sentence and a translation example. The parallel translation examples may include some auxiliary information such as grammar information representing types of parts of speech.
In the example-based approach, sentences established as correct and natural expressions are used and, therefore, if an input sentence matches any parallel translation example, it is possible to obtain a correct and natural translation result.
On the other hand, since the number of parallel translation examples is limited, a translation example that matches the input sentence cannot always be found. In such a case, translation is impossible using this approach only. Therefore, if there is no parallel translation that perfectly matches the input sentence, a process is required to find a parallel translation not perfectly matching but similar to the input sentence, and to address the difference between the parallel translation and the input sentence using some means. Such a process, however, possibly leads to lower translation accuracy.
In the following, processing in a translation apparatus taking the example-based approach will be described in detail.
In the example-based translation apparatus, a plurality of example sentences in Japanese (hereinafter referred to as Japanese Examples) and English translations of the Japanese Examples (hereinafter referred to as English Examples) are stored in correspondence with each other.
In the translation apparatus as such, if a Japanese sentence is to be translated to English, a matching process between the Japanese sentence and the Japanese Examples is executed, so that one Japanese Example is selected from the plurality of Japanese Examples. Thereafter, in the translation apparatus, an English Example corresponding to the selected Japanese Example is read from the storage unit.
It is not always the case that a Japanese sentence input by the user perfectly matches the selected Japanese Example.
Therefore, in the translation apparatus described above, the selected Japanese Example and the English Example corresponding to the Japanese Example are used as Examples as a base (hereinafter referred to as Base Examples) for translating the input sentence, and a unique process, which is different apparatus by apparatus, is executed on unmatched portions.
Japanese Patent Laying-Open No. 2006-012168 (Patent Document 1) discloses one such unique process.
According to the method described in this reference, if a portion of the input sentence differing from a corresponding portion of the Japanese Example is identified, said portion of the input sentence is designated as an input unmatched portion, and the corresponding portion of the Japanese Example is designated as a Base Example unmatched portion.
Thereafter, according to the method, the input unmatched portion is used to select a set of subsidiary example sentences (a set of Japanese and English Examples), and possible translation candidates corresponding to the input unmatched portion are determined from the set of example sentences. By this method, according to a prescribed selection algorithm, one candidate is selected from among the candidates. The selected translation candidate is used to replace the Base Example unmatched portion of the translation Base Examples.
Japanese Patent Laying-Open No. 2002-278964 discloses a translation supporting apparatus including a (rule-based) machine translation apparatus, an (example-based) translation memory, and a similar sentence translation unit.
In the translation supporting apparatus, when translation request is made to the translation memory, the similar sentence translation unit searches in a parallel translation database for a similar sentence using the input sentence. Further, the similar sentence translation unit passes the input sentence to the machine translation apparatus. Then, the similar sentence translation unit obtains the result of analysis of the input sentence from the machine translation apparatus, and compares the result of analysis of the input sentence with the similar sentence.
Thereafter, based on the result of comparison, the similar sentence translation unit transmits a word/phrase of the input sentence differing from the similar sentence to the machine translation apparatus, and obtains a parallel translation word/phrase of the differing word/phrase. Then, the similar sentence translation unit replaces the word/phrase differing from the input sentence in the parallel translation of the similar sentence with the obtained parallel translation word/phrase. As a result, in the translation supporting apparatus, differing portions in the parallel translation results are automatically modified through coordination with the machine translation apparatus and, hence, it is unnecessary for the user to make corrections.    Patent Document 1: Japanese Patent Laying-Open No. 2006-012168    Patent Document 2: Japanese Patent Laying-Open No. 2002-278964